Hacker News Boosts Ghost Pepper’s Case for Fully Local Speech-to-Text on macOS

The Show HN thread for Ghost Pepper made a simple argument resonate with developers: if voice input is going to sit inside coding and writing loops, the lowest-friction version may be the one that never leaves the laptop. The HN post described Ghost Pepper as a 100% local macOS speech-to-text utility that records while Control is held, then transcribes and pastes on release. That premise, plus 440 points and 191 comments, was enough to push it beyond a small utility launch and into the broader conversation about private, agent-adjacent desktop tooling.

The GitHub README is unusually concrete about the stack. Ghost Pepper targets macOS 14.0+ on Apple Silicon, runs as a menu-bar app, and downloads its models locally. For speech it supports Whisper tiny.en, Whisper small.en, multilingual Whisper small, and Parakeet v3 via FluidAudio. For cleanup it uses local Qwen 3.5 models through LLM.swift to remove filler words and smooth self-corrections before pasting text into whatever field is active. The repo also states that transcriptions are not logged to disk and that debug logs stay in memory only until the app quits.

That combination is why the HN discussion mattered. A lot of AI voice products still route audio through remote APIs or hide the exact boundary between local and cloud processing. Ghost Pepper instead leans into a transparent architecture: WhisperKit for transcription, LLM.swift for cleanup, Hugging Face-hosted downloads, and a conventional macOS permission model for microphone and Accessibility access. It reads less like a venture-backed assistant and more like an open desktop utility assembled from the current local-model ecosystem.

Why HN responded

• The app turns a familiar global-hotkey interaction into an on-device AI workflow instead of a cloud transcription subscription.
• The supported model list makes the tradeoffs explicit: faster English-only options, larger multilingual options, and several cleanup-model sizes.
• The privacy claim is not just marketing language; the README spells out that models run locally and that transcriptions are not written to disk.

Ghost Pepper is not trying to be a general-purpose agent. But the HN reaction suggests that the market for agent-era tools also includes narrower utilities that solve one input problem well. As more developers mix keyboard, voice, and local models in the same workflow, projects like Ghost Pepper look like practical building blocks rather than side demos.